Multi-Intent Natural Language Understanding Framework for Automotive Applications: A Heterogeneous Parallel Approach

Abstract

Natural language understanding (NLU) is an important aspect of achieving human–machine interactions in the automotive application field, consisting of two core subtasks, multiple-intent detection, and slot filling (ID-SF). However, existing joint multiple ID-SF tasks in the Chinese automotive domain face two challenges: (1) There is a limited availability of Chinese multi-intent corpus data for research purposes in the automotive domain; (2) In the current models, the interaction between intent detection and slot filling is often unidirectional, which ultimately leads to inadequate accuracy in intent detection. A novel multi-intent parallel interactive framework based on heterogeneous graphs for the automotive applications field (Auto-HPIF) was proposed to overcome these issues. Its improvements mainly include three aspects: firstly, the incorporation of the Chinese bidirectional encoder representations from transformers (BERT) language model and Gaussian prior attention mechanism allow each word to acquire more comprehensive contextual information; secondly, the establishment of a heterogeneous graph parallel interactive network efficiently exploits intent and slot information, facilitating mutual guidance; lastly, the application of the cross-entropy loss function to the multi-intent classification task enhances the model’s robustness and adaptability. Additionally, a Chinese automotive multi-intent dataset (CADS) comprising 13,100 Chinese utterances, seven types of slots, and thirty types of intents were collected and annotated. The proposed framework model demonstrates significant improvements across various datasets. On the Chinese automotive multi-intent dataset (CADS), the model achieves an overall accuracy of 87.94%, marking a notable 2.07% enhancement over the previous best baseline. Additionally, the model performs commendably on two publicly available datasets. Specifically, it showcases a 3.0% increase in overall accuracy on the MixATIS dataset and a 0.7% improvement on the MixSNIPS dataset. These findings showcase the efficacy and generalizability of the proposed model in tackling the complexity of joint multiple ID-SF tasks within the Chinese automotive domain.